task: centralize frame-key serialization #3489

Summary

The PyTorch video-analysis pipeline currently hand-computes frame-key padding
widths in multiple places and stores per-frame outputs under string keys such as
frame0, frame09, or frame00042.

This works, but it is a low-level serialization detail that is easy to get
wrong, e.g. see #3485 where where log10(0) for empty frames raised an
OverflowError.

Motivation

• Frame-key formatting is an implementation detail, but callers currently need
  to reason about key_str_width.
• Width calculations are subtle because they depend on the highest frame index,
  not just the total number of frames in an intuitive way.
• Read/write code should ideally share one canonical frame-key formatter.

Design Goals

• Centralize all frame-key logic in one place.
• Preserve the existing on-disk key format for backward compatibility.
• Reduce duplicate logic across pickle and shelf code paths.